Download The Power of CLV: Managing Customer Lifetime Value at IBM

informs
Vol. 27, No. 4, July–August 2008, pp. 585–599
issn 0732-2399 eissn 1526-548X 08 2704 0585
®
doi 10.1287/mksc.1070.0319
© 2008 INFORMS
Practice Prize Report
The Power of CLV: Managing Customer
Lifetime Value at IBM
V. Kumar
J. Mack Robinson College of Business, Georgia State University, Atlanta, Georgia 30303,
[email protected]
Rajkumar Venkatesan
Darden Graduate School of Business, University of Virginia, Charlottesville, Virginia 22904,
[email protected]
Tim Bohling
Americas Market Intelligence, IBM Corporation, New York, New York 10589, [email protected]
Denise Beckmann
Americas Market Intelligence, IBM Corporation, Atlanta, Georgia 30327, [email protected]
C
ustomer management activities at firms involve making consistent decisions over time, about: (a) which
customers to select for targeting, (b) determining the level of resources to be allocated to the selected
customers, and (c) selecting customers to be nurtured to increase future profitability. Measurement of customer
profitability and a deep understanding of the link between firm actions and customer profitability are critical
for ensuring the success of the above decisions. We present the case study of how IBM used customer lifetime
value (CLV) as an indicator of customer profitability and allocated marketing resources based on CLV. CLV
was used as a criterion for determining the level of marketing contacts through direct mail, telesales, e-mail,
and catalogs for each customer. In a pilot study implemented for about 35,000 customers, this approach led to
reallocation of resources for about 14% of the customers as compared to the allocation rules used previously
(which were based on past spending history). The CLV-based resource reallocation led to an increase in revenue
of about $20 million (a tenfold increase) without any changes in the level of marketing investment. Overall,
the successful implementation of the CLV-based approach resulted in increased productivity from marketing
investments. We also discuss the organizational and implementation challenges that surrounded the adoption
of CLV in this firm.
Key words: customer relationship management; customer lifetime value; field experiment; return on marketing
contacts; missing value imputation
History: This paper was received September 21, 2006, and was with the authors 3 months for 2 revisions;
processed by John Roberts. Published online in Articles in Advance May 7, 2008.
1.
IBM Background
time of this study, the channels used for contacting
a customer depended largely on the past relationship
with a customer. For example, customers who have
employees ranging from 100 to 999 (the midmarket)
are contacted primarily through direct mail, telesales,
e-mails, and catalogs.
Each year, IBM sorts customers based on their score
on a customer-level metric and prioritizes marketing
contacts to customers based on this score. The choice
of the metric (used to score customers) is considered
as one of the primary factors that determines the
return on marketing and is constantly refined by IBM
with the objective of maximizing potential value from
customers. Through the 1990s, a customer spending
score (CSS) was used to score customers. CSS was
defined as the total revenue that can be expected from
IBM is one of the leading multinational hightechnology firms that markets hardware, software,
and services to B2B customers. IBM intends to proactively manage individual customer relationships to
maximize overall firm profitability. A variety of marketing factors including product/service innovation,
product/service pricing, transaction channel, mass
market advertising, and individual customer contacts
are expected to impact customer profitability. Among
these factors, the level of customer contacts is most
applicable for customization across customers, and is
the focus of differential resource allocation for managing customer profitability at IBM. Customers are
contacted through several channels, such as salesperson, direct mail, telesales, e-mail, and catalogs. At the
585
586
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585–599, © 2008 INFORMS
a customer in the next year. Customers from the top
one or two CSS deciles, depending on the customer
segments (small to medium businesses or large companies), are selected for future targeting.
Over the years, additional components were added
to augment CSS. Ultimately, the CSS score was
abandoned because it focused primarily on customer
revenues (i.e., the top line) and largely ignored the
variable cost of serving a customer—and therefore,
the bottom line. Trials of customer profitability and
customer lifetime value (CLV) were proposed as an
alternative for scoring customers. In this study, we
present the case study of how CLV was measured,
evaluated, and implemented in IBM for prioritizing marketing resources among customers in the
midmarket.
The objectives of the customer selection process at
IBM are summarized by the following questions:
• “Which customers should be selected for targeting?”
• “Is there a way to determine the level of
resources to be allocated to those customers?”
• “How can the selected customers be nurtured in
order to increase future profitability?”
In trials we tested a simple management belief:
“Can an increase in contacts to the right customers
create high value from low-value customers when
all other drivers are similar?” To accomplish these
objectives, IBM adopted the following CLV management framework. The CLV management framework
is intended to guide the marketing activity directed
towards customers each year.
2.
Customer Lifetime Value
Management Framework and
Impact
The measurement of CLV and the alignment of
marketing resources to customers with the highest
potential value are at the core of the proposed CLV
management framework illustrated in Table 1.
Although the components of this framework have
been proposed separately in past literature (Gupta
and Zeithaml 2006), this is the first study to provide an integrated framework that is also suitable for
field implementation. In this paper, we discuss implementation, through a field study, of the CLV management framework for a sample of IBM’s customers
(the midmarket), and the impact of this implementation on marketing productivity. Although field studies
are relatively scarce in the literature, they are essential for establishing the external validity of marketing
strategies and vastly improve the adoption of the proposed strategy by firms across industries. The field
study that we report here also illustrates the interplay
Table 1
Customer Lifetime Value Management Framework
Process
Measure customer lifetime
value (CLV)
Identify the drivers of CLV
Determine optimal level of
contacts for each customer
that would maximize their
respective CLV
Develop propensity models to
predict what product(s) a
customer is likely to purchase
Reallocate marketing contacts
from low CLV customers to
high CLV customers.
Purpose
To obtain a measure of the potential
value of IBM customers.
So that managers can influence the CLV
To guide managers on the level of
investment required for each
customer
To develop a product message when
contacting a customer
To maximize marketing productivity
between the short-term tactical decisions such as marketing contacts and the long-term strategic focus on
CLV. In addition, this paper discusses some initiatives
at IBM that further expand upon the foundations of
the field study reported here.
The results from the field study are encouraging.
Marketing resources were reallocated based on CLV
calculations for 14% of the customers. The resource
reallocation involved contacting a certain set of customers in 2005 who were not contacted until 2004,
but who are predicted to have high CLV. On average, the revenue from these customers increased tenfold. The total increase in revenues in the test sample
for the midmarket customers, as compared to previous years, is about $20 million dollars. This increase in
revenue is obtained without any changes in the level
of marketing investment.
The rest of the article is organized as follows. In §3,
we provide an overview of how the customer profitability management framework was implemented
in IBM. Section 4 explains the context in which the
study was conducted, and §5 provides details on
the first stage, model development and measurement.
The successful application of the customer profitability management framework to IBM’s customer base
has been remarkable, and has been described, along
with some extensions, in §6. In §7, we discuss the
organizational challenges most companies would face
when implementing the framework. The paper concludes with a summary in which further issues such
as transferability to other industries and the limitation
of the customer profitability management framework
are discussed.
3.
Assessing the Impact of the
Customer Lifetime Value
Management Framework
We now describe the two-stage process used to
determine whether the CLV management framework
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585–599, © 2008 INFORMS
Figure 1
587
Implementing Customer Lifetime Value Management Framework
Purchase history
Marketing information
Gross margin data
Computation of customer selection metrics
Compare customer selection metrics in
choosing the valuable customer.
“Who to Target”
Estimation
time period
Identifying the optimal contact strategy for
customer selection using CLV
Developing ContactContact strategy
“How many times to ContactContact”
Propensity modeling
“What products to Pitch”
Field study
Experimentation
time period
Contact group
No contact group
Compare
performance
achieved its objectives. In the first stage, several
models were developed to generate inputs for implementation. In the second stage, a field study was conducted based on recommendations from the models
developed in the first stage. In the field experiment,
we tested an initial theoretical assumption within IBM
that, an increase in contacts to the right customer creates
high value from low value customers when all other drivers
are similar.
3.1. First Stage: Model Development
IBM used CLV, coupled with other qualitative inputs
from the sales team, to base their marketing decisions
for a sample of midmarket customers. Figure 1 illustrates the phased approach that was implemented,
and the various steps that were executed in order are
explained below:
Phase I: Defined and developed a model to measure CLV for each customer.
Phase II: Conducted a prediction exercise to compare the performance of customers rank ordered based
on the traditionally used metrics with that of CLV.
Phase III: To maximize CLV, we developed an optimal contact strategy to allocate contact resources to
each customer. Once the guidelines for the optimal
number of contacts (in different channels of communication—telephone, catalog, e-mail, and direct mail)
with the customers are established, we then observed
the performance in the marketplace based on our
recommendations.
Phase IV: Propensity models were built for each
product category to identify the product to feature in
addition to the CLV measure, which helps to select the
customers for targeting; and the optimization process,
which suggests the contact strategy.
3.2. Second Stage: Implementation
Phase V: Customers were split into two groups—
(1) not contacted so far (Not Contacted Until 2004
group), and (2) previously contacted (Contacted by
2004 group). Marketing contacts were then reallocated
to align resources to the high CLV customers.
Phase VI: Customers with potential for higher CLV,
but not contacted so far—i.e., the Not Contacted until
2004 group—were contacted in 2005 as per the recommendations of the CLV-based approach. The performance of this group of customers was compared
between the years 2004 and 2005 to illustrate the
impact of the model recommendations. Further,
the model recommendations were evaluated even for
the intersection of the Contacted by 2004 group and the
Contacted in 2005 group to see if we missed out on
existing source of revenues. In other words, if we recommended that someone be contacted in 2005, did
the customers provide higher revenue than the customers who were not contacted?
4.
Study Context—Data
Data on midmarket companies were used to evaluate the impact of the CLV-based framework. The midmarket customers represent companies with number
588
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585–599, © 2008 INFORMS
of employees in the range 100–999 at the enterprise
level and with total enterprise revenues to IBM in
the last three years (2001–2003) that were greater than
$25,000. The total number of enterprises in the data
set was greater than 20,000. Because some enterprises
can have more than one establishment (e.g., different
locations making independent decisions), we conduct
our analysis at the establishment level. We aggregate
all the customer transactions over a month and conduct our analyses at the monthly level because at
least a month is required to sort out the complexities
involved in selling to other businesses, and most of
the transactions registered within a certain month in
the customer database correspond to the same customer order. We have a total of 2.5 million observations (72 months each for the 35,131 establishments)
in the analysis group.
where,
5.
Computation of CLV requires predictions on three
aspects: (a) the level of marketing contacts directed
towards customer i in time period j (“MT” in Equation (1)), (b) the probability that a customer would
purchase in each time period (“pBuy” in Equation (1)), and (c) the contribution (in $s) provided by
the customer in each time period (“CM” in Equation (1)).
Model Likelihood. The three aspects involved in the
computation of CLV are inherently correlated. The
level of marketing contacts directed towards a customer depends on customer characteristics, past customer behavior, and the past level of marketing
resources allocated towards the customer. The probability that a customer would purchase is likely to
be dependent on the level of marketing resources
directed towards the customer, and finally, the customer provides profits to the firm only if they
purchase. In our model framework we allow for
correlations among these aspects of firm and customer behavior. We model marketing contacts, probability of purchase, and contribution margin jointly
through a “seemingly unrelated regression”-based
model structure. The likelihood that summarizes our
model structure is provided below,
First-Stage Model Development
and Measurement
5.1. Phase I: Measurement of CLV
We adopt the always-a-share approach for measuring
CLV because it is more appropriate for the noncontractual setting of IBM (Reinartz and Kumar 2000,
Rust et al. 2004, Venkatesan and Kumar 2004). The
always-a-share approach assumes that there is only
dormancy in a customer-firm relationship and that
customers never terminate their relationship with a
firm. This assumption allows for a customer to return
to purchasing from a firm after a temporary dormancy, and when the customer returns to the relationship they retain the memory about their prior relationship with the firm.
We define the lifetime value for customer i as the
net present value of cash flows they provide over
a three-year period (36 months). Theoretically, CLV
models should estimate the value of a customer over
the customer’s lifetime. However, in many firms,
including IBM, three years is considered to be a good
estimate for the horizon over which the current business environment (with regard to technology, competition, etc.) would not change substantially. Hence,
most customer relationship management (CRM) decisions are made based on CLV estimates over a rolling
three-year window. Further, in most cases the majority of a customer’s lifetime value is captured within
the first three years (Gupta and Lehmann 2005). For
example, if the retention rate is equal to 75%, and
the discount rate is equal to 20%, then three years
accounts for about 86% of the CLV.1 We therefore
measure CLV as,
T
+36 pBuy = 1 · CM
ij · MC
ij MT
ij
−
(1)
CLVi =
j−T
1 + r
1 + rj−T
j=T +1
1
We thank the area editor for providing us with this example.
CLVi = Lifetime value for customer i,
pBuyij = Predicted probability that customer i will
purchase in time period j,
ij = Predicted contribution margin provided
CM
by customer i, in time period j,
ij = Predicted level of marketing contacts (or
MT
touches) directed towards customer i in
time period j,
MC = Average cost for a single marketing contact, this is assumed to be $7 in the study,
j = Index for time periods; months in this
case,
T = Marks the end of the calibration or observation time frame, and
r = Monthly discount factor; 0.0125 in this
case (amounts to a 15% annual rate).
LMT Buy CM
∝
T
N i=1 j=1
PrBuyij∗ ≤ 0 MTij 1−Buyij
· PrCM∗ij = CMij Buyij∗ > 0 MTij Buyij
(2)
where
Buyij∗ = customer i’s latent utility for purchasing in
time period j.
Please refer to Appendix A for further details on
the likelihood. As indicated in Table 2 and explained
in Appendix A, the proposed framework extends the
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585–599, © 2008 INFORMS
Table 2
589
Comparison of Model Contributions
Studies
Reinartz et al. (2005)
Venkatesan and Kumar (2004)
Lewis (2004)
Fader et al. (2005)
Donkers et al. (2007)
Current study
Imputing missing values
in contribution margin
Joint estimation
of CLV components
Modeling
firm’s decisions
Field
experiment
Forward-looking cost
allocation strategy
No
No
No
No
No
Yes
No
No
No
Yes
Yes
Yes
No
No
Yes
No
No
Yes
No
No
No
No
No
Yes
No
No
No
No
No
Yes
previous individual level, and always-a-share-based
CLV models along four critical dimensions: (a) modeling firm decisions, (b) providing a forward-looking
cost allocation strategy, (c) imputing missing contribution margin, and (d) accommodating for unobserved
dependence among levels of marketing contacts, purchase incidence, and contribution margin.
We used the first 54 months of data to estimate
model parameters, and the CLV score was computed
for each customer over the next 36 months (2002
through 2004), as described in Equation (1). Further
details on model estimation and model comparisons
are provided in the Technical Appendix, which is
available at http://mktsci.journal.informs.org.
5.1.1. Drivers of CLV. Although firms are interested in knowing the lifetime value of their customers,
they are also keen on identifying the factors that are in
their control that could increase the value of their customers. Reinartz and Kumar (2003) identified the factors that explain the variation in the profitable lifetime
duration among customers. The antecedents of profitable lifetime duration are grouped as exchange characteristics and customer heterogeneity. The exchange
characteristics define and describe the nature of
the customer-firm exchange, whereas firmographic
variables capture customer heterogeneity. Different
exchange characteristics that we include as drivers of
purchase propensity and contribution margin include
past customer-spending level, cross-buying behavior,
purchase frequency, recency of purchase, past purchase activity, and the marketing contacts by the firm.
We also evaluated interactions between the drivers to
accommodate for nonlinearity in the influence of the
drivers on purchase incidence and contribution margin. To accommodate for the diminishing returns to
marketing efforts (Venkatesan and Kumar 2004), we
use the logarithm of the marketing contacts variables
as predictors in the purchase incidence and contribution margin equations.2
2
We also evaluated both a linear and a quadratic term of marketing
contacts as independent variables in the model. We found that the
in-sample fit and predictive accuracy were better when log of marketing contacts was used, given the range of marketing contacts in
our data.
Drivers of marketing contacts were based both on
theory and discussions about how marketing contacts
are executed at IBM. The drivers of contact history
include past customer spending, past levels of marketing contacts, customer relationship (or exchange)
characteristics such as cross buying and recency, and
past purchase activity. The past levels of marketing
contacts were by definition equal to zero for establishment without contact history. Similar to the purchase propensity and contribution margin equations,
we included interactions between the drivers in the
marketing contacts equation. When assigning predictors, we ensured that there is at least one unique
predictor for each dependent variable, i.e., marketing
contacts, purchase propensity, and contribution margin, to ensure model identification (Greene 1993).
Customer firmographics were included as drivers
of a customer-specific intercept ( in Equation (A7) in
Appendix A) in all three components of our model
framework: marketing contacts, purchase propensity,
and contribution margin. The various firmographics
included were the sales of an establishment (a measure of the size of the establishment), an indicator
for whether the establishment belonged to the B2B or
B2C industry category, and the installed level of PCs
in the establishment (PcQ), a measure of the level of
demand for IT products in the establishment.
The coefficient estimates of the drivers of CLV are
reported in Table 3. The reported values are the posterior means and variances. A parameter is considered
not significant if a zero exists within the 2.5th percentile and 97.5th percentile values of the posterior
distribution for that parameter.
Regarding the level of marketing contacts, the
results reflect IBM practices related to contacting customers. The level of marketing contacts for a customer depends on the recent purchase behavior of the
customer, which is captured through the covariates—
lagged indicators of purchase incidence, lagged contribution margin, and the lagged number of purchases
made by a customer so far. Whereas a customer’s past
purchase behavior in general determines whether a
customer is contacted, the specific level of marketing
contacts directed towards a customer in a particular
month is influenced to a large extent by the level contacts for the customer in the two prior months. Finally,
590
Table 3
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585–599, © 2008 INFORMS
Estimation Results
Coefficients∗
Independent variables
Level of marketing contacts
Lagged level of contacts
Two-period lagged level of contacts
Lagged average number of purchases
Two-period lagged indicator of purchase
Interaction of cross buying and recency
Lagged contribution margin
Purchase incidence
Lagged indicator of purchase
Two-period lagged indicator of purchase
Lagged average level of contribution margin
Log of lagged level of contacts
Interaction of cross buying and recency
Interaction of log of lagged level of contacts
and lagged indicator of purchase
Contribution margin
Lagged contribution margin
Lagged average contribution margin
Cross buying
Frequency of purchases
Log of lagged level of contacts
Interaction of cross buying and recency
Mean
Variance
07366
03239
05836
47114
−00149
06016
00316
00333
02158
14023
00035
01128
06573
02172
00056
00041
−00047
00004
00902
00891
00026
00012
00074
00002
08612
07442
02858
73692
0079
−0038
00247
00325
01075
1887
00105
00565
∗
Mean and variance are computed using the 5th through 95th percentiles
of the posterior sample.
the interaction of cross buying and recency of purchase reflects IBM’s strategy of focusing marketing
activities on customers who have been actively purchasing products across several categories.
The significant positive influence of the two lagged
indicators of purchase indicates the existence of stationarity in customer purchase incidence. Also, customers who have spent more (as captured by the
average level of contribution margin), and customers
who have made a recent purchase (indicating that the
customer is active) and have purchased across a wider
range of product categories (as captured by cross buying) are more likely to purchase in the current month.
Finally, marketing contacts have a significant positive
influence on customer purchase incidence, and the
influence of marketing contacts is enhanced for customers who have made a recent purchase.
Similar to purchase incidence, past customer spending seems to have a positive reinforcing effect on
spending in the future. Over and beyond the lagged
effects of customer spending, customers who have a
greater breadth of relationship (i.e., cross buying) and
customers who have purchased frequently in the past
provide a higher contribution margin. Also, the contribution margin potential is further enhanced for customers who have both a greater relationship breadth
and have made a recent purchase.
The heterogeneity distribution for the customerspecific intercepts (provided in Technical Appendix
W1) indicate that although IBM allocates more marketing contacts for customers who have a higher sales,
the purchase incidence and contribution margin is
lower for these customers. This is possible because
customers who have higher sales (or larger companies) in general split their purchases across several
vendors (Bowman and Narayandas 2004). Similarly,
customers in the B2B industry are allocated more marketing contacts, but these customers have a lower purchase incidence and lower contribution margin than
customers in the B2C industries. The coefficients for
customer sales and industry category imply that IBM
may benefit from reallocating the marketing contacts.
Finally, customers who have a large installed base of
PCs (an indicator of the demand for IT-related products) have a higher purchase incidence and contribution margin and are contacted more by IBM.
5.2.
Phase II: Comparing Traditional Customer
Selection Metrics with CLV
Difference Between CLV and the Traditionally Used
Metrics. Although recency-frequency-monetary value
(RFM), past customer value (PCV), and CSS are commonly used for computing the customer’s future
value, they suffer from the following drawbacks.3
RFM and PCV are not forward looking and do
not consider whether a customer is going to be
active in the future. These measures consider only
the observed purchase behavior and assume that
past behavior reflects future behavior. RFM assumes
that the recency, frequency, and monetary value of a
customer’s purchase explain the future value of the
customer. It fails to account for other factors (e.g.,
marketing actions) that help to predict the customer’s
future purchase behavior and the customer’s worth
to the firm. Also, the weights given for R, F, and M
greatly influence the computation of a customer’s
worth. PCV fails to account for factors (e.g., cross
buying) influencing future purchase behavior of customers. It also does not incorporate the expected cost
of maintaining the customer in the future. This limits its use as a valuable input in designing customerlevel marketing strategies. As mentioned before, CSS
focuses only on customer revenue and ignores the
cost of serving a customer. On the other hand, the
CLV measure incorporates the probability of a customer being active in the future, the future contribution margin, and the marketing costs to be spent to
retain the customer. All these factors used to measure
CLV are essential for designing customer-level marketing strategies that maximize firm value.
3
Details on the measurement of the traditional metrics are provided
in the Technical Appendix at http://mktsci.journal.informs.org.
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585–599, © 2008 INFORMS
Table 4
591
A Comparison of Metrics for Customer Selection
Using the first 54 months of data to predict the next 18 months of purchase behavior
Percent of cohort
(selected from top)
15
Average revenue
Gross value
Variable costs
Net value
Customer lifetime
value
CSS
RFM
Past customer
value
30427
9184
107
9077
21789
6659
114
6544
22622
6966
110
6856
23542
7185
104
7081
Notes. The reported values are in dollars (expressed as a multiple of the actual numbers) per customer and are cell
medians. The net result was identifying the top customers who provided the best customer value.
Metrics Evaluation. The CLV score for each customer was computed using information from the predictions of marketing contacts, purchase incidence,
and contribution margin (obtained from the proposed
modeling framework illustrated in Appendix A), as
well as the unit marketing costs for each channel.
Seventy-two months of historical data were available
for model development. Traditional metrics were also
computed using the first 54 months of data. Customers were then rank ordered based on the CLV
measure as well as on the traditionally used metrics.
The comparative performance of the customers (i.e.,
the observed profits provided by the customers in the
last 18 months) in the top 15% of each metrics’ list
clearly shows the power of CLV to identify the best
customers for future targeting (see Table 4). Previous
research in contractual settings has found that current profit is a good indicator of future profitability
(Donkers et al. 2007). In contrast, our results indicate
that in noncontractual settings, at least with regard to
selecting high-potential customers for future targeting, current profit performs worse than estimates of
future profitability.
5.3.
Phase III: Optimal Contact Strategy
Forward-Looking Cost Allocation Strategy. In
Phase III, we use a genetic algorithm to obtain
the “optimal contact strategy” for each customer
(Venkatesan et al. 2007). Once the posterior distribution of the parameter estimates of the model used to
measure CLV (Phase I) are obtained, we vary the frequency of marketing contacts for each customer and
then calculate the sum of the expected CLV of all the
customers in the sample. For each customer, we calculate a CLV corresponding to each sampled value of
the posterior distribution of the parameter estimates
for that customer, and the average of the CLVs across
the posterior distribution for the customer provides
the expected CLV. The optimization algorithm maximizes the sum of expected CLVs for all the customers.
To compute the optimal contact strategy, we assume
that IBM would make the same number of contacts
to a particular customer every month over the threeyear prediction horizon. IBM contacts its midmarket
customers through direct mail, telesales, and e-mail.
Within IBM, the implementation of marketing contacts within each channel is managed by separate
groups. Obtaining the support and participation of
each marketing channel group would have significantly delayed the implementation of the field experiment. Also, the field experiment was intended to
illustrate the usefulness of the CLV management
framework within the organization. We expected the
success of the field experiment to improve the chances
of organization-wide adoption of this framework and
also to help us obtain the participation of all the marketing channel groups in future campaigns. Therefore,
we identify the optimal level of marketing contacts
across all channels, and not the optimal level for each
channel.
Given the posterior sample of model coefficients,
the optimization algorithm is implemented by varying the level of lagged marketing contacts and twoperiod lagged marketing contacts for each customer.
For the first time period in the prediction horizon (j =
T + 1, in Equation (1)), we first compute the level of
marketing contacts (Equation (A1) in Appendix A).
Given the predicted level of marketing contacts, we
then predict purchase incidence (Equations (A2) and
(A3) in Appendix A), and predict contribution margin (Equation (A4) in Appendix A) given the level of
marketing contacts and purchase incidence. The predicted quantities for the first time period are then substituted back as independent variables to predict the
values in the second time period and so forth. These
predictions are carried forward to 36 months to calculate CLV from Equation (1). This process is repeated
for each sampled value in the posterior distribution
of the parameters to compute the expected CLV for
each customer for a given level of marketing contacts.
Different values of lagged marketing contacts and
two-period lagged marketing contacts that are simulated from the optimization algorithm will lead
to different predictions of marketing contacts, purchase incidence, and contribution margin, and hence
expected CLV. The objective of the optimization algorithm is to find the optimal level of marketing contacts for each customer that would maximize the
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
592
Marketing Science 27(4), pp. 585–599, © 2008 INFORMS
sum of expected CLVs of all the customers. The optimization algorithm therefore maximizes the sum of
expected CLVs by varying 10,000 parameters (i.e.,
varying the lagged and two-period lagged levels of
marketing contacts for the 5,000 customers in the estimation sample). We set the parameters in the genetic
algorithm as follows: population size = 200, probability of crossover = 08, probability of mutation = 025,
and convergence criteria = difference in optimal solution over the last 10,000 iterations is less than 0.1%.
Further details on the genetic algorithm used in this
study are provided in online Appendix W3. Note that
even though we estimate a single marketing contact
response parameter for all the customers, the optimal
level of marketing contacts that would maximize a
customer’s expected CLV is different because the values of the other covariates in the model vary by each
customer.
The output from the optimal resource allocation
model produced input into the decision-making process about the number of contacts in each channel for each customer in various customer segments
described below. The differences in suggested optimal contact frequencies across various customer segments are shown below. At the time of the study, the
firm was using CSS as a key metric for targeting customers and allocation of contact resources. The segments themselves are fixed. For example, one segment
could be recently acquired customers, and the other
segment could be dormant customers, and so on. Figure 2 shows the contact frequencies for each of these
Figure 2
A Comparison of Contact Strategies
RECOMMENDED-Using CLV
140
120
5.4.
Average
100
80
60
40
20
0
A
B
C
D
E
F
Customer segments
STATUS QUO-Using customer spending score
140
120
Average
100
80
60
40
20
0
segments (whose names are not revealed due to confidentiality reasons). However, as shown in Table 5,
when we used the CLV metric, the contact frequencies changed (as a result of maximizing CLV) for
these segments (quite the contrary to the CSS metric).
CLV, coupled with other techniques, provided further
insights for IBM to redistribute marketing contacts by
customer segments.
For example, some customers in Segment D are not
being contacted at all, and some other customers in
that segment are contacted at a very low frequency.
The proposed framework recommends that customers
in Segment D be contacted more so that the resulting
CLV is higher. This exercise indicates that the resource
allocation strategy suggested by CLV is different from
that suggested by a CSS metric. However, we would
be able to assess whether the different resource allocation strategy suggested translates into higher profits
for the firm only through a field experiment.
The optimal level of marketing contacts is used to
guide the determination and budgeting of customerlevel marketing resources. However, recognizing the
fact that the optimization framework does not consider competitive reactions and changes in the product/service space, the optimal level of contacts served
only as a planning tool. During implementation, a
customer was contacted until a purchase, or when
an additional contact would result in negative profitability. Because the customer responses in the current period are augmented to the customer database,
repetition of the CLV measurements and the determination of the corresponding optimal level of contacts over time are expected to reduce the discrepancy
between the planned optimal level of resources and
the realized level of resources.
A
B
C
D
Customer segments
E
F
Phase IV: Prediction of Purchase Probabilities
for Each Product Category
The CLV ranking determines which customers to target; the propensity model determines which products to pitch to the targeted customers. In Phase IV
we estimated a customer’s propensity to purchase in
each of the product categories. A series of binomial
logit models are estimated (one for each product category) to predict customer purchase propensity. The
use of sophisticated modeling approaches that accommodate for coincidence of product category purchases
such as the multivariate probit model (Seetharaman
et al. 2006) are also being evaluated currently for
predicting purchase propensity. However, the simple binary logistic model approach was used in this
study because it was the primary methodology used
in the organization at the time, and we wanted to
clearly evaluate the benefits from implementing the
CLV approach first.
Propensity models were built to predict the propensity of buying products within the following
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585–599, © 2008 INFORMS
Table 5
593
Optimization Strategies
Customer spending score (CSS)
Customer lifetime value
Low
High
High
Direct mail/Telesales/Catalog/Email:
Current interval is 4.82 days
Optimal interval is 1.9 days
Gross value:
Current value is $10,936
Optimal value is $17,809
Direct mail/Telesales/Catalog/Email:
Current interval is 6.3 days
Optimal interval is 5.3 days
Gross value:
Current value is $53,488
Optimal value is $90,522
Low
Direct mail/Telesales/Catalog/Email:
Current interval is 9.7 days
Optimal interval is 12.6 days
Gross value:
Current value is $743
Optimal value is $1,203
Direct mail/Telesales/Catalog/Email:
Current interval is 8.4 days
Optimal interval is 8.3 days
Gross value:
Current value is $1,091
Optimal value is $2,835
categories—hardware, software, and services. Our
assumption is that the customer’s need for certain
product types as well as familiarity with the focal categories are the key drivers of product category choice.
The drivers of product category choice that have a significant influence include (1) the proportion of samecategory purchases, i.e., the dominance of a category
over others; (2) the depth of same-category purchases
measured as the number of products purchased within
the focal category, i.e., the knowledge of the focal category; and (3) the breadth of same-category purchases
measured as the number of different product types
purchased within the focal category. Marketing contacts are not included as a predictor in the logistic
regressions because the customer database currently
does not register the product message in a marketing contact. The percentage of correct classifications
for the propensity models were 88%, 84%, and 89%
for the hardware, software, and services categories,
respectively. A product message was used in a marketing contact to a customer if the predicted purchase
propensity for the category was greater than 0.5 for
that customer. The predictions from the propensity
models were used to provide further insight on the right
product/service message to convey when contacting the
selected customers.
6.
Second Stage—Implementation
The second stage involves the field implementation
of the models developed and validated for measuring CLV, and the selection of the product message in
the first stage. The general guidelines for the second
stage, which involves the implementation of the CLV
management framework, are provided below:
1. Assess the distribution of customer value.
2. Identify customers with high potential value that
have been ignored in the past.
3. Select customers for contacting in the next year
by dropping customers that were targeted in the past
but are predicted to have low CLV, and picking customers who are identified in Step 2.
4. Provide the new customer list to the salespeople.
Make them aware of the optimal level of resources
and the product message identified for each customer
in Phase III (in the first stage) as a guideline.
5. Record the revenue obtained from each customer
and the resources that were required to obtain the
revenues.4
6. Going forward to the subsequent year, reestimate the CLV and propensity models developed in
the first stage with the new data obtained the prior
year.
We now provide details about the implementation
of the above implementation guidelines in our study.
6.1. Phase V: Resource Reallocation
In Phase V, the marketing resources are aligned with
customers who have the highest potential. The customers are first divided into two groups—customers
who have not been contacted at all (the Not Contacted Until 2004 group), and customers who were
contacted previously (the Contacted by 2004 group). In
each group, the customers were further divided into
deciles, and the mean CLV in each decile is provided
in Table 6.
Table 6 shows that customers in Decile 10 for the
Contacted by 2004 group have a negative mean CLV,
and the customers in Decile 1 for the Not Contacted
Until 2004 group have a higher CLV than Decile 2
(and onwards) of the Contacted by 2004 group.
4
Salespeople also use their judgment and intuition when deciding
on the resources required for each customer. Further, salespeople
from the multiple product groups coordinate their sales calls as
deemed appropriate.
Marketing Science 27(4), pp. 585–599, © 2008 INFORMS
Resource Reallocation Based on CLV∗
Figure 3
Not contacted
until 2004 ($)
Contacted
by 2004 ($)
Customer segment
1
350,471
2,124,483
Super high CLV
2
993
125,460
High CLV
3
4
5
6
7
669
638
623
611
534
43,681
23,624
17,499
13,675
10,513
Medium CLV
8
9
10
444
369
80
8,051
5,023
(35)
Low CLV
Decile
∗
The values reported here have been adjusted by a constant factor of the
actual figures.
Based on the distribution of CLVs across the deciles,
it was recommended that marketing contacts be reallocated from customers in Decile 10 in the Contacted
by 2004 group to customers in the top three deciles
of the Not Contacted Until 2004 group. For the customers in the top three deciles of the Not Contacted
Until 2004 group, we then obtained the predictions
for purchase probability for the three product categories from Phase IV. Customers who had a high
purchase propensity for at least one of the three
product categories were selected as candidates for
resource reallocation. The candidate customers were
then rank ordered based on their CLV (provided in
Table 6). Marketing resources were allocated based
on this rank order (i.e., higher CLV candidate customers were first allocated resources) until all the
resources that were available from Decile 10 of the
“Contacted by 2004” group were exhausted. The level
of resources allocated to each candidate customer was
determined based on the optimal contact strategy
described in Phase III. This process for resource reallocation resulted in some customers in Deciles 1 and all
the customers in Deciles 2 and 3 in the Not Contacted
Until 2004 group being allocated marketing resources
for 2005.
6.2. Phase VI: Field Study
Recommendations from Phase V coupled with inputs
from other groups, including the sales managers,
formed the basis for the marketing contact strategy
implemented in 2005.5 The process of implementing the strategy involved contacting the customers
who were not contacted until 2004 but who exhibit
the potential for high CLV in 2005—i.e., Deciles 1
through 3 in the Not Contacted Until 2004 group.
5
2005 can also be considered as a “true holdout” time period.
Measuring the Impact of the Contact Strategy
Average revenue/customer
(for the same group of customers)
Percent of establishments
w/purchase
$10X (2005)
1.6Y% (2005)
Y% (2004)
%
Table 6
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Revenue
594
$X (2004)
“No Contact until 2004”
“Contacted in 2005”
6.2.1. Impact. The results reveal that on average
the revenue from the customers who were not contacted until 2004, but contacted in 2005 increased by
10 times (across all customers), as shown in Figure 3.
To ensure that the increase in revenue is due to contacts and pitching the right products/services, the
revenues for the customer group who requested not
to be contacted at all—Do Not Contact group—were
compared for the years 2004 and 2005. The average
revenue for this group in both years showed similar revenues (i.e., similar to the revenue provided in
2004 by the “not contacted until 2004 but contacted
in 2005” group). We therefore infer that the higher
performance in 2005 for the group of customers who
were not contacted in 2004 was due to contacting the
right customers with the right message. As explained
before, the selection of the customers and the corresponding messages were based on the models developed in Phases 1 through 5, thereby illustrating the
economic impact of these models. The effectiveness
of the propensity models was reflected in the superior performance of the sales revenue metric. In other
words, if inappropriate products were pitched, then
the firm would not have realized higher revenues.
Further, the average number of contacts (8.9) per customer across the customers was close to the predicted
number of contacts (8.1) for this group.
The net result was higher revenues (although it
is still not the net profit, given that certain selling
[general and administration] expenses have to be
accounted for, the comparison holds good), and
higher value in 2005 versus 2004 for the no contact until 2004 but contacted in 2005 group of customers (Deciles 1 through 3 in the Not Contacted Until
2004 group). The total increase in revenues for the
no contact until 2004 but contacted in 2005 group
of customers is about $19.2 million dollars. We use
the increase in revenue among the Contacted by 2004
group of customers from 2004 ($751 M) to 2005 ($793
M), which is about 5%, as the year-to-year growth
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585–599, © 2008 INFORMS
Table 7
Further Analysis of the “Not Contacted Until 2004” Group
Dormant customers
reactivated in 2005
Number of customers
Change in revenue from 2004 ($)
Average increase in revenue
per customer ($)
273
114 M
4176 M
Customers
active in
2004 and 2005
1827
7.6 M
4160
in revenue from IBM customers. After adjusting for
the annual growth in customer revenue, the incremental revenue due to resource reallocation among
the No Contact until 2004 group of customers is about
$19.1 M. The direct marketing contact cost for contacting these establishments is about $95,200 (given the
average unit cost of $7 per contact). The model development costs for this study are estimated to be about
$25,000. The ROI for IBM in this study is therefore
around 160.6
Similar to Elsner et al. (2004), we investigate the
sources of incremental value. The incremental value
in 2005 for the No Contact Until 2004 but Contacted in
2005 group of customers is derived from two sources.
Table 7 shows that $7.6 million (approximately 40%) is
obtained from the increase in purchase amount from
customers who were active (or purchased) in 2004
and $11.4 million (approximately 60%) is obtained
from new purchases from customers who did not
purchase in 2004, i.e., reactivation of dormant customers. About 273 customers who were dormant in
2004 were reactivated in 2005; therefore, the average
increase in revenue from reactivating dormant customers was about $41,758. One thousand eight hundred twenty-seven customers were active in both 2004
and 2005. Therefore, the average increase in revenue
from existing customers was $4,160. The resource
reallocation therefore seems to result in both growing existing customers and reactivating dormant customers, with slightly more emphasis on reactivating
dormant customers.
The total revenue from the customers who were
contacted in 2004 and 2005 (based on our model recommendations) was over 750 million dollars (after
accounting for the direct marketing expenses). Therefore, our model identified both existing and new
sources of revenue. The reallocation of marketing
resources led to about a 3% increase in profits from
2004 to 2005 for the midmarket customers from contacting only 1% of the IBM customers. The impact
of the CLV-based campaign on the profitability of
midmarket customers has accelerated the adoption
of the CLV management framework for other customer segments in IBM. One such initiative that is
6
The ROI measure excludes any selling (general and administration) expenses.
595
notable is the implementation of the CLV management framework for marketing multiple service solutions to very large customers (i.e., customers with
greater than 1,000 employees), the large enterprise
segment. Within this initiative, sales team contacts are
prioritized based on the potential value of the customers, and predictions from propensity models for
the various service options act as inputs for the message of the sales team.
7.
Organizational and Implementation
Challenges
The reorientation required to implement a CLV management framework has to be carefully managed by
the firm. In this section, we outline some of the
challenges faced and the steps that are necessary
to manage this change successfully. Customers are
more central to firms than brands and brand equity,
although current management practices in many firms
do not reflect this shift (Rust et al. 2004). Traditionally, in many companies, marketing and sales activities for individual product groups are implemented
by a product-specific sales team, sometimes with little
idea of what activities have been planned by the other
product groups targeted at the same customer.
Although many multiproduct firms may strive to
coordinate their messaging strategy, they fall short in
the customer’s mind when the firm’s various marketing and sales groups set forth group-specific communication goals that often are not in sync with
each other or with the overall strategy. Customers
often feel they are dealing with multiple companies because they receive uncoordinated communications from both sales and marketing. This is
possible because some customers have been observed
to have a high affinity for the firm’s products. On the
other hand, some potential high-value customers get
ignored. Applying the CLV management framework
helps a firm to understand which products, specific
customers, and/or customer segments are likely to
buy at varying time frames. Thus, to reap the benefits
of the CLV management framework, a firm should
try to develop strategies and tactics from a customer’s
perspective rather than from a product perspective.
This transformation presents multiple organizational
and implementation challenges. We summarize some
of these challenges next.
Organizational challenges can be broadly categorized along business and people dimensions. The
business dimension requires defining and articulating the business case for change and the desired
outcomes of change. Quantification of the projected
return on investment is required not just within specific business units but also across multiple key stakeholders and across business units. Cross-organization
596
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
collaboration is essential to identify and assess key
stakeholders, degree of risk, and cost of change in
the planning and implementation phases. As to the
human element, managers and professionals who
have traditionally been responsible for all aspects of
the marketing of a single brand now have to be
empowered with the ability and responsibility across
brands, products, and units. Thus, the following steps
are recommended:
1. Generate awareness of the need to change.
2. Create a desire to participate and support the
change through innovative rewards and incentives.
3. Disseminate knowledge of how to change.
4. Empower ability to implement the change on a
day-to-day basis.
5. Provide managerial reinforcement to keep the
change in place.
Success of campaigns implemented based on the
CLV management framework depends largely on
understanding customer-level campaign responses
and documenting the relationship progression to provide effective feedback to the system. The three critical aspects for campaign implementation are outlined
next.
Applying the Time-Based Probability Measure. The
measure for the probability to purchase in each product category and customer value for each quarter or
any other time frame of interest is provided in the
campaign execution output file for each customer. The
recommendation from the output file may suggest a
significant change in the customer contact pattern.
Figure 2, described earlier, illustrates the actual versus recommended contact pattern (aggregated across
customers within each segment, over three product
categories across four quarters) for each of the six different customer segments. The recommendation runs
counter to the current practice. The analysis suggests
that customer segments C, D, and E, which were
being contacted least frequently, should be contacted
most frequently to maximize CLV across various marketing initiatives. If, after several contacts, the customer does not purchase a product in the expected
quarter, it is prudent to reevaluate the contact frequency with the customer in the next quarter and
onwards for that product category.
Documentation of Progress. To effectively assess performance, the purchases made by the customer need
to be recorded in real time. For the customers who
have already made the purchase, the product category purchased needs to be documented, and further contact efforts need to be tailored based on this
information.
Performance Tracking and Monitoring. To provide the
inputs into the CLV computation, measures such
as marketing communications cost, manufacturing
Marketing Science 27(4), pp. 585–599, © 2008 INFORMS
cost, service delivery cost, retention and acquisition costs, and revenues need to be monitored and
reported. Other metrics that guide performance are
often reported to the management. Examples of these
are cross-sell ratio, win-back ratio, share of wallet,
increase in profits, and overall return on investment.
8.
Conclusions and Future Steps
The CLV management framework proposed can allow
a firm to refine and improve its customer contact
strategy. Using the CLV-based approach, firms can:
• Increase the return on their marketing investment by allocating resources towards customers who
are most likely to provide value in the next year.
• Identify products to sell as bundles.
• Reallocate the excess resources (after targeting
the most likely customers to buy in a given time
period) to other prospects (acquiring new customers
or reactivating dormant customers).
Plans are in place to measure the CLV for the larger
set of customers (greater than 1,000 employees) and
focus on even higher marketing productivity to
expand this concept to the entire organization. Two
major initiatives that are being considered based on
the CLV management framework include (a) segmentation and profiling, and (b) understanding customer
migration.
Segmenting and Profiling. The segmentation and
profiling initiative is intended to guide the acquisition activities. Customers are segmented into distinct
groups based on their CLV, share/size of wallet estimates and firmographics-based profiles of these customer segments are built. The expected CLV and the
segment affiliation of the CLV deciles for the midmarket customers discussed in this study are presented
in Table 6. The customer deciles are grouped to form
four segments based on the CLV; Super High CLV,
High CLV, Medium CLV, and Low CLV. For these
customer segments the firmographics used to build
the customer profiles primarily include the industry
category, the number of employees, the sales of the
establishment, and the installed level of PCs in the
establishment. Only those firmogaphics that had discernible differences across CLV segments are included
for profile analyses. The customer acquisition contacts
are then directed towards prospects that match the
profiles of the Super High CLV and the High CLV
segments.
Understand Customer Migration. The customer migration initiative harnesses CLV measurements over
time to improve retention activities. Customer migration is defined as the movement of customers from
either the high-value segments (Super High CLV and
High CLV segments in Table 7) in a year to the lowvalue segment (Low CLV segment in Table 7) in the
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585–599, © 2008 INFORMS
next year, or vice versa. Separate propensity models are built to identify (1) the drivers of customer
migration to the low-profitability segments and (2)
the drivers of customer migration to high-profitability
segments. The various drivers that are considered
include the level and change in marketing resource
allocation, the rate of acquisition of new products by
the customer, the intensity of financing availed by
the customer, and the changes in the depth of purchasing exhibited by a customer. The drivers of customer migration are used to guide customer retention
strategies.
We expect that the framework presented here has
immediate application in industries where a firm has
direct contact with its customers, such as online retailing, telecommunications, hospitality, and direct-toconsumer financial services firms. In industries where
firms market through resellers or agents, firms may
adopt a multistage profitability analysis where the
profitability of the intermediary (i.e., the resellers or
the agents) is determined by the profitability of the
end consumers associated with the intermediaries.
Marketing resource allocations can then be determined based on the intermediary’s profitability.
Limitations and Future Research. The CLV management framework presented here has limitations that
could serve as fruitful avenues for future research.
The model does not include the impact of competitive marketing efforts on CLV. Although we did
accommodate for competitive effects indirectly by
the imputation of unobserved contribution margin
values, a richer understanding of customer responsiveness to marketing contacts can be obtained by
including the level of marketing contacts initiated
by competitors at the same time. Accurate data on
competitive initiatives is of course hard to obtain,
but the customer transaction information can be augmented with information from periodic surveys of
customers on the intensity of competitor actions and
the customer’s attitudes to IBM’s products (Keller and
Lehmann 2006). The CLV management framework
presented in this study is similar in some aspects
to previous research in sales call planning (Lodish
1971, Zoltners and Sinha 2005). For example, both
our framework and CALLPLAN (Lodish 1971) consider cost and revenue, as well as customer responsiveness, to optimize across customer allocation of
marketing resources. Whereas CALLPLAN estimates
customer responsiveness based on salesperson estimates, our model framework estimates them based on
historical data on customer revenue and salescall frequency. Also, our framework includes a component
on predicting customer’s product choices, which is
absent in the CALLPLAN framework. Future research
should evaluate the relative superiority of the proposed framework with earlier sales-planning frameworks such as CALLPLAN. The design of our field
597
experiment does not allow us to estimate the contribution of accurately predicting customer’s product
purchase propensities (Phase IV) towards the growth
in profitability for the “Not Contacted in 2004” group.
Research indicates that appropriate cross-selling models can improve the profitability of sales campaigns
significantly (Kumar et al. 2008). Future research that
compares the contribution provided by better models
that accurately select profitable customers and models that accurately identify the product message (such
as the cross-selling models) would provide a valuable contribution to the marketing productivity literature. Although we measured CLV and determined the
customers to target based on CLV at the individual
customer level, we did not provide any recommendations about optimal level of marketing resources
for each customer. As also suggested by Rust and
Chung (2006), studies that document the benefits of
customer-level optimal marketing resource allocation
based on CLV would provide a good contribution to
the literature.
Our framework does not include the impact of
macroeconomic trends and new product introductions by competitors on CLV because the current
study dealt with a single instance of CLV measurement. These are factors that are important to consider
when the CLV measurement is carried over multiple years and would prove especially critical for
understanding customer migration. The interaction
of customer appreciation initiatives such as loyalty
programs, and the frequency of communication in
each marketing channel, with customer profitabilitybased marketing resource allocation, is a fruitful
area for future research. Finally, the development of
sophisticated models based on the multivariate probit
framework for predicting customer’s basket choices
in the context of the CLV management framework is
a good avenue for future research.
In summary, the impact on marketing productivity realized in the field study has been instrumental
in the adoption of the customer profitability management as a core framework for IBM’s marketing strategies for customer acquisition, retention, growth, and
win back.
Acknowledgments
The authors thank the editor, John Roberts, Gary Lilien, and
the members of the INFORMS Practice Prize Committee for
their encouragement and support of this study.
Appendix A. Details on Model Framework for
Measuring CLV
As mentioned in the text, measurement of CLV requires
predictions of (a) marketing contacts (MT) directed towards
a customer, (b) the purchase propensity of the customer
[pBuy], and (c) the contribution margin provided by the
customer given purchase (CM).
598
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585–599, © 2008 INFORMS
We first model the log of the level of marketing contacts
allocated by the firm towards customer i as
T
1 + u1ij logMTij = 1i + x1ij
(A1)
where x1ij , 1 , 1i , u1ij are a vector of predictor variables,
a vector of corresponding coefficients, an individual-level
intercept, and an error term.
We assume that customer i, purchases from the firm
only when the customer’s latent utility for purchasing from
the firm (Buyij∗ ), exceeds a certain threshold, set to zero in
this case. We do not observe the latent utility of the customer, but observe a binary outcome variable regarding
whether the customer purchased or did not purchase in
time period j. The latent utility is mapped to the binary
outcome variable (Buyij ) as follows:
Buyij∗ > 0
if Buyij = 1
Buyij∗ ≤ 0
if Buyij = 0
(A2)
The latent utility, (Buyij∗ ), for customer i to purchase from
the firm in time period j is then modeled as a function of
predictor variables in a linear model
T
2 + u2ij
Buyij∗ = 2i + x2ij
(A3)
where, similar to Equation (A1), x2ij , 2 , 2i , u2ij are a vector
of predictor variables, a vector of corresponding coefficients,
an individual-level intercept, and an error term. Finally, we
assume that a latent variable, CM∗ij , represents the amount
spent by customer i in time period j, irrespective of whether
it is with the firm, as a function of predictor variables with
a linear structure
T
3 + u3ij
CM∗ij = 3i + x3ij
(A4)
where, similar to Equations (A1) and (A3), x3ij , 3 , 3i , u3ij
are a vector of predictor variables, a vector of corresponding
coefficients, an individual-level intercept, and an error term.
If the customer purchased from the firm in time period j,
then the firm observes the contribution margin provided by
the customer, i.e.,
CMij = CM∗ij if Buyij = 1
= unobserved
if Buyij = 0
(A5)
When data about customer transactions with all the firms
in a product category are available (e.g., in scanner panel
data), we can treat the contribution margin provided by a
customer when there is no-purchase as equal to zero. However, in CRM databases, there is rich information about
customer transactions with a single firm, but almost no
information about the customer’s transactions with other
firms. Therefore, when we don’t observe a purchase from a
customer in time period j, we cannot assume that the customer did not make a purchase at all. A no-purchase in a
CRM database would only imply that the customer has not
made a purchase with the focal firm, but may have a purchase with a competitor. We therefore treat the time periods when a customer does not provide any revenue to the
firm as “missing data.” We then impute these missing values
based as random realizations from a normally distributed
prior distribution for the missing contribution margins. We
assume the mean and variance of the prior distribution
to be the empirical mean and variance in the contribution
margin provided by a customer to IBM in a calibration
sample.
For observations with no purchase incidence (i.e., Buyij
equals zero), there is a possibility that the customer has
made a purchase with a competitor. However, the imputation of the missing contribution margins should also take
into account a customer’s interpurchase time. We therefore
compute a customer’s average interpurchase time from the
calibration sample used to compute the mean and variance for the prior distribution. The imputation of contribution margins is performed only for those observations
whose time since last purchase (or recency) is greater than
the average interpurchase time calculated in the calibration sample. For example, if a customer’s average interpurchase time in the calibration sample is three, then
only those missing contribution margin observations whose
recency is greater than three are imputed from the prior
distribution; the contribution margin is allowed to be zero
for the nonpurchase occasions whose recency is less than
three. We choose this approach for selecting the nonpurchase occasions that qualify for imputation based on the
assumptions about customer purchase patterns in IBM. We
specify the following prior for missing CMs in months j,
when the recency of purchase exceeds customer i’s average interpurchase time obtained from the calibration sample, CM∗ij recencyij > t̄i = T N− 0 i CM i2 CM , where
t̄i = average interpurchase time for customer i in the
calibration sample, i CM = average observed contribution margin provided by customer i in the calibration
sample, and i2 CM = variance in the observed contribution margin provided by customer i in the calibration
sample.
The imputations are obtained as a random draw from a
truncated normal distribution (whose parameters are determined in the calibration sample as explained before) with
limits − and 0. In other words, we impute negative contribution margin values (including zero) for the no purchase
incidence occasions reflecting the contribution margin lost
to competitors on these occasions. The customer-specific
mean and variance of the prior distribution accommodates
for customer-specific variations in the spending levels. The
prior distribution parameters assume that a customer has
the same spending level with IBM as well as with the competitors. We adopt this conservative assumption because
of a lack of a reliable share of wallet information about
the customers. If reliable share of wallet information were
available, then the product of a customer’s average contribution margin with IBM and the share of the customer’s
competitor purchases can be used as the mean of the
prior distribution. Given that no information is available
about customer transactions with competition, the calibration sample provides us with the best estimate (or expectation) of customer interactions with competitors. We expect
that by treating a no-purchase occasion in this manner
we would obtain better estimates of a customer’s value,
the opportunity cost of a customer and of the customer’s
responsiveness to marketing communications. The imputation of missing contribution margin is performed only
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585–599, © 2008 INFORMS
in the estimation sample. When calculating CLV, the contribution margin is set to zero for months when the predictions indicate no purchase incidence. The imputation
of missing contribution margins in the estimation sample
are hence relevant for providing better estimates of a customer’s response elasticity, which is critical for accurately
calculating CLV. Our approach here is the first step towards
accommodating the missing information about customer’s
transactions with competition. Future research that evaluates other mechanisms for addressing this issue would be
very valuable.
The above formulation is similar to the “seemingly unrelated regression” model structure, and the predictor variables in Equations (A1), (A3), and (A4) need not be the
same. We assume that the covariance structure of the errors
in Equations (A1), (A3), and (A4) may be modeled as,

  


u1ij
1 12 13
0

  


 u2ij  ∼ N  0   12 22 23  = N3 0 y (A6)

  


u3ij
0
13
23
33
affording the possibility of correlations among the three
residuals. We fix 11 to be equal to one to ensure model identification. The covariance structure of the errors accounts
for any unobserved dependence between a firm’s decision
to contact a customer (MT), a customer’s decision to purchase from the firm (Buy), and the amount the customer
spends with the firm (CM). Letting = 1 2 3 ], and
= 1 2 3 , the simultaneous equations model gives rise
to the likelihood specified in Equation (2). The customerspecific intercept terms are obtained from a multivariate normal distribution;
i ∼ MVN Zi (A7)
where,
Zi = a p × 1 vector of customer characteristics;
= a 3 × p matrix of coefficients for the customer characteristics;
= a 3 × 3 variance-covariance matrix;
p = number of customer characteristics that are used to
capture heterogeneity.
We assume diffuse and conjugate priors for the model
parameters. We assume multivariate normal priors for ,
and . Let d denote the dimension of the vector; then, the
prior specification of is given as ∼ MVN ), where
is a d dimensional column vector of zeros, and =
100Id ; Id is a d × d dimensional identity matrix. Similarly,
let d denote the dimension of ; then the prior specification of is given as ∼ MVN , where is a d
dimensional column vector of zeros, and = Id ; Id is a
d × d , dimensional identity matrix. Inverse Wishart priors
are assumed for the variance parameters. The prior specification for is given as = IW Id , where = 15
and Id is a d × d , dimensional identity matrix. The prior
specification for y is given as y = IW 15I3 N T , where I3
is a 3 × 3 dimensional identity matrix.
Further details on the data augmentation, the prior specifications, and the Markov chain Monte Carlo (MCMC)
estimation algorithm can be obtained from Cowles et al.
(1996).
599
References
Bowman, D., D. Narayandas. 2004. Linking customer management
effort to customer profitability in business marketing. J. Marketing Res. 41(4) 433–447.
Cowles, M. K., B. P. Carlin, J. E. Connett. 1996. Bayesian tobit
modeling of longitudinal ordinal clinical trial compliance data
with nonignorable missingness. J. Amer. Statist. Assoc. 91(433)
86–98.
Donkers, B., P. C. Verhoef, M. G. de Jong. 2007. Modeling CLV:
A test of competing models in the insurance industry. Quant.
Marketing Econom. Forthcoming.
Elsner, R., M. Krafft, A. Huchzermeier. 2004. Optimizing Rhenania’s
direct marketing business through dynamic multilevel modeling (DMLM) in a multicatalog-brand environment. Marketing
Sci. 23(2) 192–206.
Fader, P. S., B. G. S. Hardie, K. L. Lee. 2005. RFM and CLV: Using
iso-value curves for customer base analysis. J. Marketing Res.
42(4) 415–430.
Greene, W. H. 1993. Econometric Analysis, 3rd ed. Prentice Hall, Inc.,
Upper Saddle River, NJ.
Gupta, S., D. Lehmann. 2005. Managing Customers as Investments:
The Strategic Value of Customers in the Long Run. Wharton School
Publishing, Upper Saddle River, NJ.
Gupta, S., V. Zeithaml. 2006. Customer metrics and their impact on
financial performance. Marketing Sci. 25(6) 718–739.
Keller, K. L., D. R. Lehmann. 2006. Brands and branding: Research
findings and future priorities. Marketing Sci. 25(6) 740–759.
Kumar, V., R. Venkatesan, W. Reinartz. 2008. Performance implications of adopting a customer-focused sales campaign. J. Marketing. Forthcoming.
Lewis, M. 2004. The influence of loyalty programs and short-term
promotions on customer retention. J. Marketing Res. 41(August)
281–292.
Lodish, L. M. 1971. CALLPLAN: An interactive salesman’s call
planning system. Management Sci. 18(4) 25–40.
Reinartz, W. J., V. Kumar. 2000. On the profitability of long lifetime customers: An empirical investigation and implications
for marketing. J. Marketing 64(4) 17–32.
Reinartz, W. J., V. Kumar. 2003. The impact of customer relationship
characteristics on profitable lifetime duration. J. Marketing 67(1)
77–99.
Reinartz, W., J. Thomas, V. Kumar. 2005. Balancing acquisition and
retention resources to maximize customer profitability. J. Marketing 69(January) 63–79.
Rust, R., T. S. Chung. 2006. Marketing models of service and relationships. Marketing Sci. 25(6) 560–580.
Rust, R., K. N. Lemon, V. A. Zeithaml. 2004. Return on marketing:
Using customer equity to focus marketing strategy. J. Marketing
68(January) 109–127.
Seetharaman, P. B., S. Chib, A. Ainslie, P. Boatwright, T. Chen,
S. Gupta, N. Mehta, V. Rao, A. Strijnev. 2006. Models of multicategory choice behavior. Marketing Lett. 16(3–4) 239–254.
Venkatesan, R., V. Kumar. 2004. A customer lifetime value based
framework for customer selection and resource allocation strategy. J. Marketing 68(October) 106–125.
Venkatesan, R., V. Kumar, T. Bohling. 2007. Optimal CRM using
bayesian decision theory: An application to customer selection.
J. Marketing Res. 44(November) 579–594.
Zoltners, A., P. Sinha. 2005. Sales territory design: Thirty years of
modeling and implementation. Marketing Sci. 24(3) 313–331.